Getting Started with Deep Learning Using Python

In the first chapter, we had a very close look at deep learning and how it is related to machine learning and artificial intelligence. In this chapter, we are going to delve deeper into this topic. We will start off by learning about what sits at the heart of deep learning—namely, neural networks and their fundamental components, such as neurons, activation units, backpropagation, and so on.

Note that this chapter is not going to be too math heavy, but at the same time, we are not going to cut short the most important formulas that are fundamental to the world of neural networks. For a more math-heavy study of the subject, readers are encouraged to read the book Deep Learning (deeplearningbook.org) by Goodfellow et al.

The following is an overview of what we are going to cover in this chapter:

• A whirlwind tour of neural networks...

Demystifying neural networks

Let's start this section by finding the answers to the question, “Why are neural networks called 'neural'?”. What is the significance behind this term?

Our intuition says that it has something to do with our brains, which is correct, but only partially. Before we get to the reason why it is only partially correct, we need to have some familiarity with the structure of a brain. For this purpose, let's look at the anatomy of our own brains.

A human brain is composed of approximately 10 billion neurons, each connected to about 10,000 other neurons, which gives it a network-like structure. The inputs to the neurons are called dendrites and the outputs are called axons. The body of a neuron is called a soma. So, on a high level, a particular soma is connected to another soma. The word "neural" comes from the word...

Different types of neural network

So far, we have learned what feed-forward neural networks look like and how techniques such as backpropagation and gradient descent are applied to it in order to optimize their training process. The binary classification problem we studied earlier appears to be too naive and too impractical, doesn't it?

Well, there are many problems that a simple NN model can solve well. But as the complexity of the problem increases, improvements to the basic NN model become necessary. These complex problems include object detection, object classification, image-caption generation, sentiment analysis, fake-news classification, sequence generation, speech translation, and so on. For problems like these, a basic NN model is not sufficient. It needs some architectural improvements so that it can solve these problems. In this section, we are going to study two...

Exploring Jupyter Notebooks

While working on a project relating to deep learning, you must deal with a huge amount of variables of various types and arrays of various dimensions. Also, since the data contained in them is massive and keeps changing after nearly every step, we need a tool that helps us to observe the output produced by each step so that we can proceed accordingly. A Jupyter Notebook is one such tool. Jupyter Notebooks are known for their simplicity, and their wide support of features and platforms are currently the standard tool for developing deep-learning solutions. The reasons for their popularity can be understood by considering the fact that several of the top tech giants offer their own version of the tool, such as Google Colaboratory and Microsoft Azure Notebooks. Moreover, the popular code-hosting website GitHub has been providing a native rendering of Jupyter...

Setting up a deep-learning-based cloud environment

Before we begin setting up a cloud-based deep learning environment, we might wonder why would we need it or how a cloud-based deep learning environment would benefit us. Deep learning requires a massive amount of mathematical calculation. At every layer of the neural network, there is a mathematical matrix undergoing multiplication with another or several other such matrices. Furthermore, every data point itself can be a vector instead of a singular entity. Now, to train over several repetitions, such a deep learning model would require a lot of time just because of the number of mathematical operations involved.

A GPU-enabled machine would be much more efficient at executing these operations because a GPU is made specifically for high-speed mathematical calculations however, GPU-enabled machines are costly and may not be affordable...

Exploring NumPy and pandas

NumPy and pandas are the backbone of nearly every data-science-related library available in the Python language. While pandas is built on top of NumPy, NumPy itself is a wrapping of Python around high-performance C code to facilitate superior mathematical computing in Python than Python itself in its pure form can provide.

Almost all deep learning software developed in Python in one way or another depends upon NumPy and pandas. It is therefore important to have a good understanding of both libraries and the features that they can provide.

NumPy

NumPy is an acronym for Numerical Python. Vanilla Python lacks the implementation of arrays, which are close analogs of the mathematical matrices used to...

Summary

We covered a lot of different things in this chapter. We started by learning the basics of a neural network and then we gradually proceeded. We learned the two most powerful types of neural networks used today—CNNs and RNNs—and we also learned about them on a high level, but without skipping their foundational units. We learned that as the complexity in a neural network increases, it requires a lot of computational power, which standard computers may fail to cater for we saw how this problem can be overcome by configuring a deep learning development environment using two different providers—AWS and Crestle. We explored Jupyter Notebooks, a powerful tool for performing deep learning tasks. We learned about the usage of two very popular Python libraries—NumPy and pandas. Both of these libraries are extensively used when performing deep learning...